Defeating the Training-Inference Mismatch via Fp16

A research paper proposes using FP16 to address the training-inference mismatch in deep learning models, but lacks community discussion to gauge its reception or implications.

Snapshot generated from the HN discussion